Windows desktop: GPT-5.5 → GPT-6.1 Sol compaction fallback fails and UI preserves obsolete model error #54030

Environment

• Windows 10 Pro x64, build 19045.
• ChatGPT sign-in (no API key used).
• App About version reported by the user: 26.1007.21434, released October 8, 2026.
• Installed desktop package version subsequently reported by codex doctor: 26.1007.9446.0. These are separate version sources; their relationship has not been established.
• Bundled runtime: 0.162.0-alpha.17.2.
• Selected and persisted model: gpt-6.1-sol, reasoning effort medium.
• Observed October 14, 2026.

Problem

Two existing desktop chats cannot continue after switching from GPT-5.5 to GPT-6.1 Sol. The picker shows GPT-6.1 Sol, but the next message triggers pre-turn compaction and fails with:

The 'gpt-5.5' model is not supported when using Codex with a ChatGPT account.
Error running remote compact task: ... HTTP 400 invalid_request_error ...

The user corrected the user-level default to model = "gpt-6.1-sol", saved it, restarted the app, and rebooted Windows. The failure persisted. Read-only inspection confirms both affected chats' stored model is also gpt-6.1-sol.

Diagnostic evidence

At 2026-10-14 16:23:33 UTC (12:23:33 EDT), the affected turn's logs show:

model=gpt-6.1-sol codex.turn.reasoning_effort=medium
run_auto_compact{reason=CompHashChanged phase=PreTurn}
previous-model compaction failed; retried with current model
previous_model="gpt-5.5"
current_model="gpt-6.1-sol"
reason=CompHashChanged
implementation=ResponsesCompactionV2
outcome="failed"
fallback_error=Some(Stream("stream closed before response.completed", None))
Failed to run pre-sampling compact

The same turn also records compaction retry warnings:

compact_error=We're currently experiencing high demand, which may cause temporary errors.
compact_error=stream disconnected before completion: stream closed before response.completed
retries=2 max_retries=2

The high-demand warning is an observation, not a proven explanation for the fallback failure.

The GPT-6.1 Sol fallback received HTTP 200 from the Responses endpoint, but the stream did not complete. Request IDs from two fallback attempts:

• 286bfd09-ea41-435b-8f0a-06c720f932db
• 7c065826-a518-4554-8aca-d02a451a3a65

The chat UI/recorded final turn error still reports the original unsupported GPT-5.5 error.

A diagnostic run under the actual Windows user account confirms:

• user configuration parses and resolves to GPT-6.1 Sol;
• ChatGPT auth is configured;
• HTTP endpoint reachability and WebSocket handshake pass;
• state, log, and history database integrity checks pass.

These checks do not establish that streamed compaction requests succeed.

Reproduction observed

1. Open an existing desktop chat with previous GPT-5.5 history.
2. Select GPT-6.1 Sol.
3. Send a new message.
4. Pre-turn compaction runs for CompHashChanged.
5. Previous-model compaction fails; current-model fallback also fails.
6. The turn stops and the UI displays the obsolete model's unsupported error.
7. Retrying after correcting the default and rebooting produces the same failure.

No minimal synthetic reproduction has been run.

Expected behavior

A chat switched to a supported model should recover through compaction. If fallback fails, the final error should identify the current-model failure as well as the original failure, so users are not led to repeatedly change an already-correct model setting.

Please advise on a supported way to recover the existing histories without deleting turns or editing databases.

Recovery investigation

The documented app-server API successfully exported the affected chat's history. A separate app-server process could not resume it because the desktop app owns the writer lock (already has an active writer). No compaction or direct database modifications were made by that recovery attempt. Manual /compact in the owning desktop app was subsequently tested; see the result below.

Related, but different failures:

• Model switch is blocked when previous-model compaction hits HTTP 429 #32453: previous-model compaction blocked by HTTP 429.
• Amazon Bedrock: model switch compacts with the previous model when comp_hash differs; the new model can't read the compaction item and the turn hangs until SSE idle timeout #48949: previous-model compaction across compatibility hashes on Bedrock.

No transcripts, credentials, repository contents, or unfiltered logs are included in this report.

Manual-compaction result (October 14, 2026)

The user ran the desktop app's /compact command at 16:37:36 UTC (12:37:36 EDT). Local logs confirm this was op.dispatch.compact / session_task.compact, explicitly using model=gpt-6.1-sol and reasoning effort medium. It failed at 16:37:40 UTC after the two retry warnings, with:

Error running remote compact task: stream disconnected before completion: stream closed before response.completed

This verifies that manual compaction with the current model also fails, independently of the initial previous-model rejection. The reason the stream closes is still unknown. No successful recovery has been established.

Second-model manual-compaction test

The user selected GPT-6 Sol in the same affected chat and ran /compact once. Local logs confirm model=gpt-6-sol, reasoning effort medium, op.dispatch.compact and session_task.compact. The operation began at 2026-10-14 16:40:09 UTC (12:40:09 EDT) and failed at 16:40:13 UTC after two retry warnings with the same error:

Error running remote compact task: stream disconnected before completion: stream closed before response.completed

This rules out a failure limited to GPT-6.1 Sol in this chat. It does not establish whether the remaining cause is server-side, transport-related during streaming, or specific to this history. No history deletion, turn rollback, direct database edits, or configuration workarounds were applied.